AI progress is wider than the frontier chat race. This October 29 edition looks at three source-backed developments with different maturity levels: usable checkpoints, a new agent model and architecture research. They did not all launch today.
OpenBMB: documents need a pipeline
MiniCPM-OCR 1.0 is a 1.3B document model for Chinese and English text, tables, formulas and structured extraction. Its repository history starts October 21 UTC; weights were updated October 23 and documentation expanded October 27–29. Those edits are not separate model releases.
The official project combines recognition, layout detection and assembly into reading-order Markdown. It releases the project and model weights under Apache-2.0. OpenBMB reports 1.2 pages per second on an RTX 5090. Its reproduction guide still lists complete scripts, exact versions and timing boundaries as work to add. That limits reproducibility; it is not a speed promise for our hardware.
Nanbeige: a teacher for smaller coding agents
Nanbeige4.3-33B targets coding and professional tool workflows under Apache-2.0. The team describes it as a teacher for future smaller agents; the announced 3B direction is not the same as an available 3B checkpoint. Its history records a weight upload October 24 UTC and evaluation additions October 27.
The model card documents a 512K context limit and different thinking-history settings for chat and multi-turn tools. Its comparative scores are publisher-reported. We have not checked quality, memory use or launcher compatibility. Context capacity is a ceiling, not a recommendation to allocate that much on a consumer PC.
Cartesia: learn the chunks instead of fixing the tokenizer
Cartesia’s October 28 H-Net scaling report describes networks that learn how to group raw bytes inside the model. Its follow-up examines why pretraining loss can translate differently into downstream performance when input processing changes.
This is architecture research, not a new drop-in speech model or a free speed upgrade for an existing GGUF. Cartesia says it plans to release its improved design. The interesting question is whether independent reproductions retain the reported efficiency advantage under matched data and compute. We have not reproduced these experiments.
Original AiBrainWorX reporting. Sources checked October 29, 2026. No model weights were downloaded and no independent performance test was performed.
